Re: [v6 05/15] mm: don't accessed uninitialized struct pages

[Pasha Tatashin]

Hi Michal,

I've been looking through this code again, and I think your suggestion 
will work. I did not realize this iterator already exist:

for_each_free_mem_range() basically iterates through (memory && !reserved)

This is exactly what we need here. So, I will update this patch to use 
this iterator, which will simplify it.

Pasha

On 08/14/2017 09:51 AM, Pasha Tatashin wrote:
> > > mem_init()
> > >   free_all_bootmem()
> > >    free_low_memory_core_early()
> > >     for_each_reserved_mem_region()
> > >      reserve_bootmem_region()
> > >       init_reserved_page() <- if this is deferred reserved page
> > >        __init_single_pfn()
> > >         __init_single_page()
> > >
> > > So, currently, we are using the value of page->flags to figure out if 
> > > this
> > > page has been initialized while being part of deferred page, but this 
> > > is not
> > > going to work for this project, as we do not zero the memory that is 
> > > backing
> > > the struct pages, and size the value of page->flags can be anything.
> >
> > True, this is the initialization part I've missed in one of the previous
> > patches already. Would it be possible to only iterate over !reserved
> > memory blocks instead? Now that we discard all the metadata later it
> > should be quite easy to do for_each_memblock_type, no?
>
> Hi Michal,
>
> Clever suggestion to add a new iterator to go through unreserved 
> existing memory, I do not think there is this iterator available, so it 
> would need to be implemented, using similar approach to what I have done 
> with a call back.
>
> However, there is a different reason, why I took this current approach.
>
> Daniel Jordan is working on a ktask support:
> https://lkml.org/lkml/2017/7/14/666
>
> He and I discussed on how to multi-thread struct pages initialization 
> within memory nodes using ktasks. Having this callback interface makes 
> that multi-threading quiet easy, improving the boot performance further, 
> with his prototype we saw x4-6 improvements (using 4-8 threads per 
> node). Reducing the total time it takes to initialize all struct pages 
> on machines with terabytes of memory to less than one second.
>
> Pasha